ML_MODE

Description: String-based tag selecting operation mode for machine learning force fields.

Mind: This tag is only available as of VASP.6.4.0.

This tag acts as a "super tag" and selects the operation mode by selecting the defaults for all other tags. Every tag that is affected by this "super tag" can be overwritten by the user by simply specifying the value for that tag. The following options are available for this tag:

ML_MODE = train: On-the-fly training

Force predictions from the machine learning force field are used to drive the molecular dynamics (MD) simulation. However, if the error estimation performed in each time step indicates a high force error an ab initio calculation is performed instead and the collected energy, forces, and stress are used to improve the machine learning force field. There are two possible cases depending on, whether an ML_AB is present in the calculation folder or not:

No ML_AB file found:
On-the-fly training is starting from scratch. Note that at the beginning of the MD run, when there is no force field available or it is still poorly trained, ab initio calculations will happen frequently. For VASP versions prior to 6.4.0 this corresponds to ML_ISTART = 0.

ML_AB file present:

Restart on-the-fly training from existing training database. Before the MD run starts the ML_AB file (a copy of the ML_ABN from a previous training run!) is read and the ab initio data (energies, forces, and stresses) and local reference configurations it contains are used to generate an initial force field. Subsequently, the on-the-fly training MD is started. For VASP versions prior to 6.4.0 this corresponds to ML_ISTART = 1.

Tip: None of the structures in the ML_AB file need to match he POSCAR file for the current MD training run in terms of the simulation box, elements, or number of atoms. However, if the same elements appear the initial force field is used for predictions in the current MD run.

The training data contained in the ML_AB file is included in the final machine learning force field, i.e., the ML_FFN file will define a force field applicable to both the structures on the ML_AB file as well as to the current MD simulation. This means that by restarting repeatedly with ML_MODE = train, and copying the ML_ABN file from the previous run to ML_AB(!), it is possible to iteratively extend the applicability of the a machine learning force field, e.g., by exploring different temperature ranges or element compositions.

ML_MODE = select: Re-selection of local reference configurations

A new machine learning force field is generated from the ab initio data provided in the ML_AB file. The structures are read and processed one by one as if harvested in an MD simulation. In other words, the same steps are performed as in on-the-fly training but the source of the data are not actual ab initio calculations in an MD run but the series of structures available in the ML_AB file. The list of local reference configurations on the ML_AB file will be ignored. Instead a new collection of local reference configurations is determined and written to the resulting ML_ABN file.

Important: This operation mode allows to generate a VASP machine learning force field from pre-computed or external ab initio data sets. In contrast to ML_MODE = refit also the local reference configurations are selected from the entire data set. For example, an ML_AB file created manually by extracting ab initio data (energies, forces, stresses) from OUTCAR files (or even other external sources) can be processed in this mode without prior knowledge of local reference configurations (only a dummy section must be added to the ML_AB file, see its documentation). Similar to on-the-fly training this mode generates ML_FFN files and ML_ABN with local reference configurations.

A new iteration through the training structures can lead to a frequent update of the force field. This is quite time-consuming. Hence, for this mode the default value of ML_CDOUB is automatically increased from 2 to 4 which will result in a much less frequent update of the force field. This leads to much more efficient calculations while practically not changing the results.

Tip: If calculations for ML_MODE = select are too time-consuming, it is useful to increase ML_MCONF_NEW to values around 10-16. Together with ML_CDOUB = 4, this often accelerates the calculations by a factor of 2-4.

The ML_AB file may contain values for CTIFOR for each training structure. These are the thresholds used to sample that structure from the previous training. The thresholds found on the ML_AB will be re-used unless a threshold is explicitly specified in the INCAR file, by means of the ML_CTIFOR tag. In the latter case the thresholds from the ML_AB file are ignored. In case the ML_AB contains no CTIFOR information and no threshold is specified in the INCAR file, the default value for ML_CTIFOR is used.

This mode automatically sets NSW = 1 and ML_CDOUB = 4.For VASP versions prior to 6.4.0 this corresponds to ML_ISTART = 3.

ML_MODE = refit : Construct a force field from an existing ML_AB file, suited for "fast" evaluation (ML_LFAST=.TRUE.).

Sets: ML_ISTART = 4, ML_LFAST = .TRUE., NSW = 1, ML_IALGO_LINREG = 4, ML_SIGW0 = 1E-7, ML_SIGV0 = 1, and ML_EPS_LOW = 1E-11.

Similar to ML_MODE = SELECT, refitting is done based on an existing ML_AB file, but the number of local reference configurations for each species is taken from the ML_AB file. Sparsification is performed on the local reference configurations, so the resulting ML_ABN file will contain the same number or fewer local reference configurations than the ML_AB file.

By default the resulting force field is geared towards "fast" evaluation to speed up production runs (ML_LFAST = .TRUE.). This comes at the cost of not being able to evaluate Bayesian error estimates.

Warning: We strongly advise to use ML_MODE = refit if no error estimates are required during production runs.

ML_MODE = refitbayesian (deprecated): Same as ML_MODE = refit, but Bayesian regression is employed.

Sets: ML_ISTART = 4, NSW = 1, ML_IALGO_LINREG = 1, and ML_LFAST=.FALSE.

This results in lower accuracy and much slower force fields than using ML_MODE = refit and should be used with caution. On the other hand, this mode allows the generation of ML_FFN files that can calculate Bayesian error estimates.

ML_MODE = run : Force-field evaluation only.

Sets: ML_ISTART=2

A previously trained machine learning force field is read from the ML_FF file, and the MD simulation is driven with predictions from the force field only, no ab initio calculations are performed and no learning is executed. This setting is typically used when the machine learning force field is considered mature and ready for production runs.

Optionally if the force field was refitted using ML_MODE = refitbayesian, the Bayesian error estimate of the energies, forces, and stress can be computed and logged in the ML_LOGFILE by setting the output frequency of the Bayesian errors by means of the ML_IERR tag. The default is ML_IERR=0.

ML_MODE = none (default): The tag is ignored.

Warning: If any option other than the above is chosen or any of them is misspelled (be careful to write everything in upper case or lower case letters) the code will exit with an error.

Tip: The user may overwrite the default by specifying any of the machine learning tags in the INCAR file.

Related tags and articles